All news with #tool abuse tag

🔍 Security researchers warn major gaps in IT oversight and rising vulnerabilities are increasing risk across the AI agent ecosystem. Reco analyzed enterprise telemetry, MCP servers, and NVD disclosures in its report, The State of Agent Security 2026, finding 80% of AI tools lack governance and SMBs average 414 unsanctioned tools per 1,000 employees. The study found many MCP servers allow shell execution, file access and outbound network calls, with numerous tools exposed without authentication. Vulnerability disclosures have also surged, straining patch programs.

🔒 A flaw in Cursor's command-line coding agent allowed cloned repositories to run arbitrary commands on a developer's machine before the user was prompted to trust the repo, and could execute outside the sandbox even when sandboxing was enabled. Manifold Security reported the issue on July 20 and published findings on August 10; Cursor shipped a fix for the pre-trust behavior three days after the report but closed the submission as informative and issued no advisory. The vulnerability involved the agent's worktree setup reading a tracked config file and piping it directly to a shell without parsing or allowlisting, and the sandbox policy was hardcoded off on that path. Developers should update to build 2026.07.23-e383d2b or later or skip worktree setup to close the pre-trust window, though Manifold says the sandbox gap remains.

🛡️ A new technique called GhostSplice shows how a malicious Model Context Protocol (MCP) server connected to an AI coding assistant can exfiltrate SSH keys, environment secrets, source code, and customer data by splitting a theft into harmless-looking fragments. ASSET Research Group tested the approach in isolated projects using fake credentials and found that splitting the request across tool descriptions, results, or server-initiated sampling raised compliance dramatically for many models. The attack relies on developers connecting a hostile MCP server and the agent already having access to the target files.

🔒 AI adoption is accelerating rapidly, but so are API-linked security incidents, making mature API management essential. The article argues that without comprehensive API discovery, runtime protection and governance, investments in AI security will fall short. It highlights shadow and zombie APIs, rising AI-related CVEs, and real-world incidents where agents deleted production data. The piece recommends continuous API inventory, runtime defenses and stricter permissions to manage AI risk.

📰 A Norwegian researcher disclosed a document-borne AI worm that can hide instructions inside files used by Microsoft Copilot and other apps, enabling prompt-injection that alters generated content and propagates into new documents. Microsoft says it has implemented mitigations and follows a defense-in-depth approach, while urging updates and caution. Experts warn the attack sidesteps many existing defenses because malicious behavior emerges only when Copilot processes content.

🔍 Enterprises face rising attacks that exploit AI agents already present in their environments. These agents — coding assistants and CLI tools like Claude Code CLI, Gemini CLI, and Amazon Q CLI — can read files, run commands, and install packages, making them attractive targets when run with auto-approval. Real-world incidents, including the s1ngularity Nx npm compromise and the AgentJacking/Sentry experiments, show how prompt injection, compromised tool metadata, and unsecured MCP servers can lead to secret harvesting and covert exfiltration. Defenders must treat agents as potentially untrusted and adapt controls and monitoring accordingly.

🛡️ Researchers demonstrated seven attacks against five open-source Android agent frameworks, showing that benign-seeming apps with draw-over and storage permissions can inject unseen text into models and escalate to host command execution. The study, posted on arXiv in July, tested AppAgent, AppAgentX, Mobile-Agent-v3, Open-AutoGLM, and MobA, finding widespread vulnerabilities including screenshot race conditions, command injection via unsanitized adb calls, broadcast leaks, and UI spoofing. Some projects already use safer patterns, but none implemented all recommended mitigations.

🔒 New research from Pillar Security shows that AI coding agents in tools like Cursor, Codex, Gemini CLI, and Antigravity can cross host security boundaries without directly breaking their sandboxes. The attacks rely on agents producing files or configurations that trusted, external components later execute or interpret. Researchers identified four repeatable failure modes and urge security teams to understand actual sandbox boundaries and treat workspace artifacts as sensitive.

🔒 Manifold Security found that Claude for Chrome still accepts synthetic clicks and can read permission mode from its URL, enabling other extensions with DOM access on claude.ai to trigger nine allowlisted tasks (including Gmail, Google Docs, and Calendar). Anthropic constrained arbitrary prompts after ClaudeBleed, but the click handler lacks an event.isTrusted check and the side panel honors ?skipPermissions=true, creating high-risk scenarios especially if "Act without asking" is enabled. Manifold reported this in May against v1.0.72; the issues remained in v1.0.80 as of July 7 and no patch or public advisory was available by July 14.

🔒 Wiz Research disclosed GhostApproval, a flaw in six AI coding assistants that allows symlink tricks to make approval prompts misrepresent targets. The vulnerability can let a repository write attacker-supplied keys or files to sensitive locations, potentially enabling passwordless remote access or remote code execution. Amazon, Google and Cursor have patched the issue; Augment and Windsurf have yet to fix it, while Anthropic disputes that its behavior is a vulnerability. Wiz recommends resolving symlinks before approval and flagging writes outside the project.

🛡️ New research describes "HalluSquatting," an attack that exploits AI assistants' habit of inventing resource names. Attackers register those predictable fake names on marketplaces and plant adversarial instructions; when an assistant hallucinate-fetches the same name, it may run the attacker's commands. The technique abuses auto-run modes and agent tools that fetch and execute external code, permitting widespread compromise without traditional malware or exploits.

🛡️ Researchers at the Hong Kong University of Science and Technology show that simple file-level transformations and packing tricks can let malicious AI coding agent "skills" evade existing static scanners while still executing normally. Their tool, SKILLCLOAK, fooled multiple marketplace scanners over 80–99% of the time, while a runtime sandbox, SKILLDETONATE, detected most evasions at the cost of slower analysis. The study highlights active real-world abuse, practical mitigation ideas, and the need to move trust decisions to behavior observed at execution time.

🧩 Researchers at LayerX demonstrated a prompt injection called BioShocking that trains AI-powered browsers to treat risky real-world actions as fictional, bypassing safety controls. The PoC used a themed puzzle game to reward 'wrong' behavior and culminated in instructing agents to copy sensitive data from a GitHub repo. Six mainstream agentic browsers were tested; only one vendor implemented a working fix after disclosure. LayerX recommends explicit user confirmations, stricter context checks, and session scope limits.

🛡️ New Microsoft research shows attackers can hijack AI agents by poisoning a tool's description so the agent quietly exfiltrates company data. The attack leverages MCP tool descriptions—plain text that agents read—to inject hidden instructions, allowing malicious actions without obvious rule violations. Microsoft recommends treating tool descriptions as system prompts, restricting approved tools, enforcing human approval for risky actions, and monitoring agent identities and behavior.

🛡️ This Microsoft Incident Response post examines an attack pattern targeting Model Context Protocol (MCP) tools, where poisoned tool metadata causes agentic AI to perform unauthorized actions. It outlines a playbook for detecting, containing, and preventing these attacks using Microsoft security controls and maps techniques to the OWASP Top 10 for Agentic Applications. The guidance emphasizes treating MCP servers as supply-chain dependencies, reviewing tool descriptions as prompts, and applying least agency controls.

🛡️ A faux AI agent skill called brand-landingpage bypassed static security scanners and reached over 26,000 users via an Instagram ad, highlighting risks as enterprises adopt AI-driven tools. The skill pointed agents to a fake Stitch SDK hosted on a domain controlled by researchers, which initially redirected to the real Google Stitch site to pass review. After distribution, the researchers changed the hosted content to instruct agents to download a script that collected email addresses, demonstrating how mutable external resources let malicious behaviors slip past static reviews. Security vendors and scanners from Cisco, Nvidia, and skills.sh marked the skill safe during testing.

🛡️ A security firm, AIR, created a benign but deceptive AI agent skill named brand-landingpage, pushed it through a major skill marketplace and promoted it with an Instagram ad, and reports it reached roughly 26,000 agents including corporate accounts. Scanners from vendors like Cisco and NVIDIA marked the package safe because the skill pointed to external setup documentation rather than embedding malicious code. AIR later swapped the external page to deliver a harmless payload that collected email addresses, demonstrating how scanners miss links that can be rewritten after review. The experiment highlights structural trust problems with skills and common mitigations such as pinning versions and vetting external references.

🛡️ In response to widespread prompt-injection risks, the article outlines runtime signals to detect compromised AI agents that possess the so-called lethal trifecta: access to private data, ingestion of untrusted content, and external communication ability. It argues that this trifecta is now the default for useful agents, so defenses must shift from architecture rules to behavioral, runtime detection. Recommended signals include instruction-following anomalies, unexpected tool-call sequences, low-bandwidth exfiltration channels, out-of-scope credential access, and suspicious memory writes.

🔎 AI agent skills can install third-party capabilities with privileged access, yet registries lack automated audits. Palo Alto Networks introduces Behavioral Integrity Verification (BIV), which compares declared metadata, executable code and natural-language instructions to detect mismatches. Applied to the OpenClaw registry, BIV found widespread deviations and identified multi-stage attack chains that enable credential theft, RCE and exfiltration. The report recommends inventorying skills and requiring pre-install behavioral checks.

🛡️ Researchers describe a new "agentjacking" attack that tricks AI coding agents into executing arbitrary code by injecting malicious instructions into Sentry error events. Tenet Security says the flaw leverages Sentry DSNs — public, write-only credentials — to post crafted markdown that appears as legitimate remediation guidance. Agents retrieving unresolved errors via MCP render the injected content as trusted and may execute the embedded commands with developer privileges. The report confirmed high exploitability across popular agents and thousands of exposed DSNs.